Add aya (#36521)

* initial commit

* small fix

* move stuff to image processing file

* remove stuff in validate turn and fix return tensor

* remove liquid stuff

* in the process of addressing comments

* changes to get the right tokenization

* new __init__ works

* fixing defulat std and mean

* works

* small testing scipt -- to be deleted before merge

* remove redundant code

* addressing comments

* fix inits, add docs templates

* refactor processor, switch to gotocr image processor

* remove image proc from init

* refactor to working llava-style architecture

* Change AyaVisionModel to AyaVisionForConditionalGeneration

* add tests

* fixups

* update doc

* Adding logits_to_keep explicitly in ayavision forward to enable compatibility with cohere model

* better variable names + remove code paths

* Updates to aya_vision.md

* address comments

* adding copied from

* make style and remove unused projector_hidden_act from config

* sort init

* include usage of fast image proc and proc on cuda in doc

* update checkpoint iin test processor

* update checkpoint in test processor 2

* remove test_model and update docstring

* skip failing tests

---------

Co-authored-by: Saurabh Dash <saurabh@cohere.com>
Co-authored-by: yonigozlan <yoni.gozlan@huggingface.co>

tests/generation/test_utils.py

@@ -113,7 +113,18 @@ from transformers.utils import is_sklearn_available
 
 
 # TODO: raushan remove this when VLMs start accepting input embeds
-VLM_CLASS_NAMES = ["llava", "idefics2", "idefics3", "mllama", "paligemma", "emu3", "gotocr2", "qwen2vl", "qwen2_5_vl"]
+VLM_CLASS_NAMES = [
+    "llava",
+    "idefics2",
+    "idefics3",
+    "mllama",
+    "paligemma",
+    "emu3",
+    "gotocr2",
+    "qwen2vl",
+    "qwen2_5_vl",
+    "ayavision",
+]
 
 
 class GenerationTesterMixin:

tests/models/aya_vision/test_modeling_aya_vision.py

@@ -0,0 +1,576 @@
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the PyTorch GotOcr2 model."""
+
+import unittest
+
+from parameterized import parameterized
+
+from transformers import (
+    AutoProcessor,
+    AyaVisionConfig,
+    is_torch_available,
+    is_vision_available,
+)
+from transformers.testing_utils import (
+    cleanup,
+    require_read_token,
+    require_torch,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        AyaVisionForConditionalGeneration,
+    )
+
+
+if is_vision_available():
+    pass
+
+
+class AyaVisionVisionText2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        seq_length=7,
+        vision_feature_layer=-1,
+        downsample_factor=2,
+        ignore_index=-100,
+        bos_token_id=0,
+        eos_token_id=0,
+        pad_token_id=0,
+        image_token_index=1,
+        num_channels=3,
+        image_size=64,
+        model_type="aya_vision",
+        is_training=True,
+        text_config={
+            "model_type": "cohere2",
+            "vocab_size": 99,
+            "hidden_size": 128,
+            "intermediate_size": 37,
+            "num_hidden_layers": 4,
+            "num_attention_heads": 4,
+            "output_channels": 64,
+            "hidden_act": "silu",
+            "max_position_embeddings": 512,
+            "tie_word_embeddings": True,
+            "bos_token_id": 0,
+            "eos_token_id": 0,
+            "pad_token_id": 0,
+        },
+        vision_config={
+            "model_type": "siglip_vision_model",
+            "hidden_size": 32,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "intermediate_size": 128,
+            "image_size": 64,
+            "patch_size": 8,
+            "vision_use_head": False,
+        },
+    ):
+        self.parent = parent
+        self.ignore_index = ignore_index
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.image_token_index = image_token_index
+        self.model_type = model_type
+        self.text_config = text_config
+        self.vision_config = vision_config
+        self.batch_size = batch_size
+        self.vision_feature_layer = vision_feature_layer
+        self.downsample_factor = downsample_factor
+        self.is_training = is_training
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.image_seq_length = (image_size // (vision_config["patch_size"] * downsample_factor)) ** 2
+        self.seq_length = seq_length + self.image_seq_length
+
+        self.num_hidden_layers = text_config["num_hidden_layers"]
+        self.vocab_size = text_config["vocab_size"]
+        self.hidden_size = text_config["hidden_size"]
+        self.num_attention_heads = text_config["num_attention_heads"]
+
+    def get_config(self):
+        return AyaVisionConfig(
+            text_config=self.text_config,
+            vision_config=self.vision_config,
+            model_type=self.model_type,
+            bos_token_id=self.bos_token_id,
+            eos_token_id=self.eos_token_id,
+            pad_token_id=self.pad_token_id,
+            image_token_index=self.image_token_index,
+            vision_feature_layer=self.vision_feature_layer,
+            downsample_factor=self.downsample_factor,
+        )
+
+    def prepare_config_and_inputs(self):
+        config = self.get_config()
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+
+        return config, pixel_values
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+        attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+        print("attention_mask", attention_mask.shape)
+        # input_ids[:, -1] = self.pad_token_id
+        input_ids[input_ids == self.image_token_index] = self.pad_token_id
+        input_ids[:, : self.image_seq_length] = self.image_token_index
+
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+    def create_and_check_model_fp16_forward(self, config, input_ids, pixel_values, attention_mask):
+        model = AyaVisionForConditionalGeneration(config=config)
+        model.to(torch_device)
+        model.half()
+        model.eval()
+        logits = model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            pixel_values=pixel_values,
+            return_dict=True,
+        )["logits"]
+        self.parent.assertFalse(torch.isnan(logits).any().item())
+
+    def create_and_check_model_fp16_autocast_forward(self, config, input_ids, pixel_values, attention_mask):
+        config.torch_dtype = torch.float16
+        model = AyaVisionForConditionalGeneration(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.autocast(device_type="cuda", dtype=torch.float16):
+            logits = model(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                pixel_values=pixel_values,
+                return_dict=True,
+            )["logits"]
+        self.parent.assertFalse(torch.isnan(logits).any().item())
+
+
+@require_torch
+class AyaVisionModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (AyaVisionForConditionalGeneration,) if is_torch_available() else ()
+    all_generative_model_classes = (AyaVisionForConditionalGeneration,) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {
+            "image-text-to-text": AyaVisionForConditionalGeneration,
+        }
+        if is_torch_available()
+        else {}
+    )
+    fx_compatible = False
+    test_pruning = False
+    test_torchscript = False
+    test_head_masking = False
+    _is_composite = True
+
+    def setUp(self):
+        self.model_tester = AyaVisionVisionText2TextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=AyaVisionConfig, has_text_modality=False)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    # overwrite inputs_embeds tests because we need to delete "pixel values" for LVLMs
+    def test_inputs_embeds(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+
+            input_ids = inputs["input_ids"]
+            del inputs["input_ids"]
+            del inputs["pixel_values"]
+
+            wte = model.get_input_embeddings()
+            inputs["inputs_embeds"] = wte(input_ids)
+
+            with torch.no_grad():
+                model(**inputs)
+
+    # overwrite inputs_embeds tests because we need to delete "pixel values" for LVLMs
+    # while some other models require pixel_values to be present
+    def test_inputs_embeds_matches_input_ids(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+            input_ids = inputs["input_ids"]
+            del inputs["input_ids"]
+            del inputs["pixel_values"]
+
+            inputs_embeds = model.get_input_embeddings()(input_ids)
+
+            with torch.no_grad():
+                out_ids = model(input_ids=input_ids, **inputs)[0]
+                out_embeds = model(inputs_embeds=inputs_embeds, **inputs)[0]
+            torch.testing.assert_close(out_embeds, out_ids)
+
+    @unittest.skip("Failing because of unique cache (HybridCache)")
+    def test_model_outputs_equivalence(self, **kwargs):
+        pass
+
+    @unittest.skip("Cohere2's forcefully disables sdpa due to softcapping")
+    def test_sdpa_can_dispatch_non_composite_models(self):
+        pass
+
+    @parameterized.expand([("float16",), ("bfloat16",), ("float32",)])
+    @unittest.skip("Cohere2's eager attn/sdpa attn outputs are expected to be different")
+    def test_eager_matches_sdpa_inference(self):
+        pass
+
+    @unittest.skip("Cohere2's eager attn/sdpa attn outputs are expected to be different")
+    def test_eager_matches_sdpa_generate(self):
+        pass
+
+    @parameterized.expand([("random",), ("same",)])
+    @unittest.skip("Cohere2 has HybridCache which is not compatible with assisted decoding")
+    def test_assisted_decoding_matches_greedy_search(self, assistant_type):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache which is not compatible with assisted decoding")
+    def test_prompt_lookup_decoding_matches_greedy_search(self, assistant_type):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache which is not compatible with assisted decoding")
+    def test_assisted_decoding_sample(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache which is not compatible with dola decoding")
+    def test_dola_decoding_sample(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache and doesn't support continue from past kv")
+    def test_generate_continue_from_past_key_values(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache and doesn't support low_memory generation")
+    def test_beam_search_low_memory(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache and doesn't support contrastive generation")
+    def test_contrastive_generate(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache and doesn't support contrastive generation")
+    def test_contrastive_generate_dict_outputs_use_cache(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache and doesn't support contrastive generation")
+    def test_contrastive_generate_low_memory(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache and doesn't support StaticCache. Though it could, it shouldn't support.")
+    def test_generate_with_static_cache(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache and doesn't support StaticCache. Though it could, it shouldn't support.")
+    def test_generate_from_inputs_embeds_with_static_cache(self):
+        pass
+
+    @unittest.skip("Cohere2 has HybridCache and doesn't support progressive generation using input embeds.")
+    def test_generate_continue_from_inputs_embeds(self):
+        pass
+
+    @unittest.skip("Failing because of unique cache (HybridCache)")
+    def test_multi_gpu_data_parallel_forward(self):
+        pass
+
+    @unittest.skip("Cohere2's eager attn/sdpa attn outputs are expected to be different")
+    def test_sdpa_equivalence(self):
+        pass
+
+    @unittest.skip(reason="SiglipVisionModel does not support standalone training")
+    def test_training(self):
+        pass
+
+    @unittest.skip(reason="SiglipVisionModel does not support standalone training")
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="SiglipVisionModel does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant(self):
+        pass
+
+    @unittest.skip(reason="SiglipVisionModel does not support standalone training")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        pass
+
+    @unittest.skip(reason="Siglip uses the same initialization scheme as the Flax original implementation")
+    def test_initialization(self):
+        pass
+
+    @unittest.skip(reason="Compile not yet supported because in LLava models")
+    def test_sdpa_can_compile_dynamic(self):
+        pass
+
+    @unittest.skip("FlashAttention only support fp16 and bf16 data type")
+    def test_flash_attn_2_fp32_ln(self):
+        pass
+
+    # todo: yoni - fix or improve the test
+    @unittest.skip("Difference is slightly higher than the threshold")
+    def test_batching_equivalence(self):
+        pass
+
+
+@require_read_token
+@require_torch
+class AyaVisionIntegrationTest(unittest.TestCase):
+    def setUp(self):
+        self.model_checkpoint = "CohereForAI/aya-vision-8b"
+
+    def tearDown(self):
+        cleanup(torch_device, gc_collect=True)
+
+    @slow
+    @require_torch_gpu
+    def test_small_model_integration_forward(self):
+        processor = AutoProcessor.from_pretrained(self.model_checkpoint)
+        model = AyaVisionForConditionalGeneration.from_pretrained(
+            self.model_checkpoint, device_map=torch_device, torch_dtype=torch.float16
+        )
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
+                    {"type": "text", "text": "Please describe the image explicitly."},
+                ],
+            }
+        ]
+
+        inputs = processor.apply_chat_template(
+            messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device, dtype=torch.float16)
+        # Forward
+        with torch.inference_mode():
+            output = model(**inputs)
+
+        actual_logits = output.logits[0, -1, :5].cpu()
+        print("actual_logits", actual_logits)
+        expected_logits = torch.tensor([0.4109, 0.1532, 0.8018, 2.1328, 0.5483], dtype=torch.float16)
+        self.assertTrue(
+            torch.allclose(actual_logits, expected_logits, atol=0.1),
+            f"Actual logits: {actual_logits}"
+            f"\nExpected logits: {expected_logits}"
+            f"\nDifference: {torch.abs(actual_logits - expected_logits)}",
+        )
+
+    @slow
+    @require_torch_gpu
+    def test_small_model_integration_generate_text_only(self):
+        processor = AutoProcessor.from_pretrained(self.model_checkpoint)
+        model = AyaVisionForConditionalGeneration.from_pretrained(
+            self.model_checkpoint, device_map=torch_device, torch_dtype=torch.float16
+        )
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "Write a haiku"},
+                ],
+            }
+        ]
+
+        inputs = processor.apply_chat_template(
+            messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device, dtype=torch.float16)
+        with torch.no_grad():
+            generate_ids = model.generate(**inputs, max_new_tokens=25, do_sample=False)
+            decoded_output = processor.decode(
+                generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+            )
+        print("decoded_output", decoded_output)
+        expected_output = "Whispers on the breeze,\nLeaves dance under moonlit skies,\nNature's quiet song."
+        self.assertEqual(decoded_output, expected_output)
+
+    @slow
+    @require_torch_gpu
+    def test_small_model_integration_generate_chat_template(self):
+        processor = AutoProcessor.from_pretrained(self.model_checkpoint)
+        model = AyaVisionForConditionalGeneration.from_pretrained(
+            self.model_checkpoint, device_map=torch_device, torch_dtype=torch.float16
+        )
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
+                    {"type": "text", "text": "Please describe the image explicitly."},
+                ],
+            }
+        ]
+
+        inputs = processor.apply_chat_template(
+            messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
+        ).to(torch_device, dtype=torch.float16)
+        with torch.no_grad():
+            generate_ids = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+            decoded_output = processor.decode(
+                generate_ids[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+            )
+        print("decoded_output", decoded_output)
+        expected_output = "The image depicts a cozy scene of two cats resting on a bright pink blanket. The cats,"  # fmt: skip
+        self.assertEqual(decoded_output, expected_output)
+
+    @slow
+    @require_torch_gpu
+    def test_small_model_integration_batched_generate(self):
+        processor = AutoProcessor.from_pretrained(self.model_checkpoint)
+        model = AyaVisionForConditionalGeneration.from_pretrained(
+            self.model_checkpoint, device_map=torch_device, torch_dtype=torch.float16
+        )
+        # Prepare inputs
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image", "url": "https://llava-vl.github.io/static/images/view.jpg"},
+                        {"type": "text", "text": "Write a haiku for this image"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
+                        {"type": "text", "text": "Describe this image"},
+                    ],
+                },
+            ],
+        ]
+        inputs = processor.apply_chat_template(
+            messages, padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
+        ).to(model.device, dtype=torch.float16)
+
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        # Check first output
+        decoded_output = processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+        print("decoded_output", decoded_output)
+        expected_output = "Wooden path to water,\nMountains echo in stillness,\nPeaceful forest scene."  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+        # Check second output
+        decoded_output = processor.decode(output[1, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+        print("decoded_output", decoded_output)
+        expected_output = 'This image captures a vibrant street scene in a bustling urban area, likely in an Asian city. The focal point is a'  # fmt: skip
+
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+    @slow
+    @require_torch_gpu
+    def test_small_model_integration_batched_generate_multi_image(self):
+        processor = AutoProcessor.from_pretrained(self.model_checkpoint)
+        model = AyaVisionForConditionalGeneration.from_pretrained(
+            self.model_checkpoint, device_map=torch_device, torch_dtype=torch.float16
+        )
+        # Prepare inputs
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image", "url": "https://llava-vl.github.io/static/images/view.jpg"},
+                        {"type": "text", "text": "Write a haiku for this image"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
+                        },
+                        {
+                            "type": "image",
+                            "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg",
+                        },
+                        {
+                            "type": "text",
+                            "text": "These images depict two different landmarks. Can you identify them?",
+                        },
+                    ],
+                },
+            ],
+        ]
+        inputs = processor.apply_chat_template(
+            messages, padding=True, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
+        ).to(model.device, dtype=torch.float16)
+        output = model.generate(**inputs, do_sample=False, max_new_tokens=25)
+
+        # Check first output
+        decoded_output = processor.decode(output[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+        # Batching seems to alter the output slightly, but it is also the case in the original implementation. This seems to be expected: https://github.com/huggingface/transformers/issues/23017#issuecomment-1649630232
+        expected_output = "Wooden path to water,\nMountains echo in stillness,\nPeaceful forest scene."  # fmt: skip
+        print("decoded_output", decoded_output)
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )
+
+        # Check second output
+        decoded_output = processor.decode(output[1, inputs["input_ids"].shape[1] :], skip_special_tokens=True)
+        print("decoded_output", decoded_output)
+        expected_output = "The first image showcases the Statue of Liberty, a colossal neoclassical sculpture on Liberty Island in New York Harbor. Standing at a"  # fmt: skip
+        self.assertEqual(
+            decoded_output,
+            expected_output,
+            f"Decoded output: {decoded_output}\nExpected output: {expected_output}",
+        )

tests/models/aya_vision/test_processor_aya_vision.py

@@ -0,0 +1,164 @@
+# coding=utf-8
+# Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import shutil
+import tempfile
+import unittest
+from typing import Optional
+
+from transformers import AutoProcessor, AutoTokenizer, AyaVisionProcessor
+from transformers.testing_utils import require_read_token, require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+
+if is_vision_available():
+    from transformers import GotOcr2ImageProcessor
+
+
+@require_read_token
+@require_vision
+class AyaVisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = AyaVisionProcessor
+
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        image_processor = GotOcr2ImageProcessor(
+            do_resize=True,
+            size={"height": 20, "width": 20},
+            max_patches=2,
+            do_rescale=True,
+            rescale_factor=1 / 255,
+            do_normalize=True,
+            image_mean=[0.485, 0.456, 0.406],
+            image_std=[0.229, 0.224, 0.225],
+            do_convert_rgb=True,
+        )
+        tokenizer = AutoTokenizer.from_pretrained("CohereForAI/aya-vision-8b", padding_side="left")
+        processor_kwargs = self.prepare_processor_dict()
+        processor = AyaVisionProcessor.from_pretrained(
+            "CohereForAI/aya-vision-8b",
+            image_processor=image_processor,
+            tokenizer=tokenizer,
+            **processor_kwargs,
+        )
+        processor.save_pretrained(self.tmpdirname)
+
+    def prepare_processor_dict(self):
+        return {"patch_size": 10, "img_size": 20}
+
+    def get_tokenizer(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
+
+    def get_image_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
+
+    def get_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    # todo: yoni, fix this test
+    @unittest.skip("Chat template has long system prompt")
+    def test_chat_template_accepts_processing_kwargs(self, **kwargs):
+        pass
+
+    # Override as AyaVisionProcessor needs image tokens in prompts
+    def prepare_text_inputs(self, batch_size: Optional[int] = None):
+        if batch_size is None:
+            return "lower newer <image>"
+
+        if batch_size < 1:
+            raise ValueError("batch_size must be greater than 0")
+
+        if batch_size == 1:
+            return ["lower newer <image>"]
+        return ["lower newer <image>", "<image> upper older longer string"] + ["<image> lower newer"] * (
+            batch_size - 2
+        )
+
+    @require_torch
+    def test_process_interleaved_images_videos(self):
+        processor = self.get_processor()
+
+        messages = [
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg",
+                        },
+                        {
+                            "type": "image",
+                            "url": "https://thumbs.dreamstime.com/b/golden-gate-bridge-san-francisco-purple-flowers-california-echium-candicans-36805947.jpg",
+                        },
+                        {"type": "text", "text": "What are the differences between these two images?"},
+                    ],
+                },
+            ],
+            [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "image",
+                            "url": "https://llava-vl.github.io/static/images/view.jpg",
+                        },
+                        {"type": "text", "text": "Write a haiku for this image"},
+                    ],
+                }
+            ],
+        ]
+
+        inputs_batched = processor.apply_chat_template(
+            messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+        )
+
+        # Process non batched inputs to check if the pixel_values and input_ids are reconstructed in the correct order when batched together
+        images_patches_index = 0
+        for i, message in enumerate(messages):
+            inputs = processor.apply_chat_template(
+                message,
+                add_generation_prompt=True,
+                tokenize=True,
+                return_dict=True,
+                return_tensors="pt",
+                padding=True,
+            )
+            # We slice with [-inputs["input_ids"].shape[1] :] as the input_ids are left padded
+            torch.testing.assert_close(
+                inputs["input_ids"][0], inputs_batched["input_ids"][i][-inputs["input_ids"].shape[1] :]
+            )
+            torch.testing.assert_close(
+                inputs["pixel_values"],
+                inputs_batched["pixel_values"][
+                    images_patches_index : images_patches_index + inputs["pixel_values"].shape[0]
+                ],
+            )
+            images_patches_index += inputs["pixel_values"].shape[0]